Short-term Residential Load Forecasting Based On Interpretable Deep Learning

The continuous penetration of distributed renewable energy and demand-side energy management systems in the residential sector has moved the scale of smart grid management down to the individual level.Short-term residential load forecasting is the basis of smart grid management at the individual level,and it provides a basis for the operation scheduling of household energy management systems and the formulation of individualized demand response programs.In recent years,deep learning-based forecasting models have provided more accurate forecast results for short-term residential load forecasting.However,due to the lack of interpretability,the short-term residential load forecasting model based on deep learning is usually regarded as a black box model,and it is difficult to know exactly the basis for the forecast made by the model.This creates barriers for customers to deeply understand the forecast results and respond quickly.To improve the interpretability of a deep learning-based short-term residential load forecasting model while maintaining forecasting accuracy,this thesis conducts an in-depth investigation of various time series forecasting methods and interpretability methods,and combines the time evolution characteristics of residential load data and other external data to propose a short-term residential load forecasting method based on interpretable deep learning in different scenarios.The main research content and innovation include the following four aspects:(1)Aiming at the lack of feature interpretability of forecasting methods based on deep learning,an interpretable residential load forecasting model based on automatic relevance determination network and long short-term memory(Long Short-term Memory,LSTM)neural network is proposed.First,the model has a built-in process of automatically selecting relevant input variables through the automatic correlation determination network,which intuitively presents the feature correlation of each input variable and improves the interpretability of the model.Afterwards,the model uses an advanced LSTM network to extract long-term time dependencies between inputs at different time steps,improving the accuracy of predictions.The final experimental results show that the proposed model has excellent prediction performance and can explain the model in terms of feature importance,which enhances the credibility of the predictive model.(2)Aiming at the problem that the deep learning residential load forecasting model lacks temporal interpretability and the traditional singlestep point forecasting model cannot reflect the changes in the load demand of residential customers,an interpretable short-term residential multi-step probabilistic load forecasting model based on multivariate LSTM and mixed attention mechanism is established.Multi-step probabilistic forecasting models.First,the multivariate LSTM network structure is used to extract the temporal dynamics of different input variables,and the mixed attention technique is embedded into the multivariate LSTM network to characterize variables and temporal importance.Additionally,this thesis enables multi-step probabilistic forecasting of residential loads by extending the interpretable multivariate LSTM to an encoder-decoder architecture,combined with quantile regression.Compared with traditional methods,the proposed method can more accurately estimate the uncertainty in the future load forecasting results and provide the knowledge of the model in terms of variable importance and temporal importance.(3)Aiming at the coupling problem of photovoltaic power generation and load consumption,this thesis proposes an interpretable residential net load probabilistic prediction model based on Transformer.The model can simultaneously predict the quantiles of the net load for multiple time steps in the future,and realize the interpretability of the prediction results.First,in order to optimize the utilization of the information carried by different variables,a local variable selection network is designed to automatically select relevant features among the available variables in different windows.The network provides insight into which variables drive predictions,giving feature importance explanations.Then,an interpretable generative Transformer module is proposed to receive integrated variable information from a local variable selection network.The generative Transformer module abandons the encoder-decoder structure,shortens the signal transmission path between prediction and input,and makes it easier for the model to take into account historical observation data.Additionally,the generative Transformer module employs an interpretable sparse selfattention mechanism to model dependencies across all input time steps.Finally,the model generates probabilistic forecasts for multiple time steps into the future via quantile regression.Experimental results show that the proposed method can present more accurate netload prediction intervals and provide explanations in terms of features and temporal patterns.(4)Aiming at the data privacy and security issues arising from model training,this thesis proposes an interpretable short-term residential load forecasting model based on federated learning.The federated learning model is based on the edge computing network framework.Its clients train their corresponding interpretable prediction models locally,and the cloud aggregates model parameters from different households to build a highprecision global generalization model.In terms of forecasting model,this thesis proposes a novel sparse automatic relevance determination network to achieve feature-level interpretability,and adopts an encoder-decoder structure to achieve multi-step forecasting of residential loads.In terms of federated training,this thesis studies the training method based on iterative federated clustering algorithm to alleviate the error caused by nonindependent and identical distribution of data.The experimental results show that the proposed residential load forecasting model based on iterative federated clustering algorithm can further improve the forecasting accuracy of the federated learning model.